27934 Deep clustering networks for hyperspectral remote sensing images


Hyperspectral image (HSI), which captures rich spectral information of objects within more than hundreds of bands as shown in Fig. 1, has been widely used in many applications such as defense and security, precision agriculture and climate change monitoring. Automatic analysis of hyperspectral image relies greatly on classification, as a means of identifying areas in the imaged scenes to some specific classes. Since data labeling in remote sensing applications is typically labor intensive and time-consuming, labeled data required for training the classifiers are often scarce, posing serious limitations for supervised classification. Clustering, as an unsupervised classification approach, which requires no labeled data, is thus crucial in this domain.

Fig. 1. Hyperspectral image contains more than hundreds of spectral bands.

In the task of HSI clustering, which aims to group pixels into different clusters (often correspond to land covers), traditional clustering methods including k-means, fuzzy c-means, density-based methods and spectral clustering often show limited performance in terms of clustering accuracy due to noise and large spectral variability within-cluster. While recent subspace clustering methods [1] by sparse coding achieved excellent clustering performance, their ability in extracting discriminative features is limited due to the adopted linear and shallow representation models. The latest deep clustering methods [2] [3] (see an example in Fig. 2) integrate unsupervised neural networks such as autoencoder with traditional clustering models like k-means and subspace clustering, which have achieved the state-of-the-art performance. However, they are mostly designed for the computer vision tasks such as clustering of facial images or online documents or hand written text (image level clustering). Analogous study in HSI clustering is very limited. Applying such algorithms directly to HSI clustering is not feasible and often leads to unsatisfactory results due to the completely different types of data and tasks. Thus, effective solutions are needed for deep clustering models in this domain.

Fig. 2. Deep subspace clustering networks [2], consisting of three convolutional encoder layers, one self-representation layer, and three deconvolutional decoder layers.


[1] S. Huang, H. Zhang and A. Pizurica, Semisupervised sparse subspace clustering method with a joint sparsity constraint for hyperspectral remote sensing images, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 12(3) (2019) 989–999.

[2] P. Ji, et al., Deep subspace clustering networks. arXiv preprint: 1709.02508, 2017.

[3] X. Peng, et al., Deep subspace clustering. IEEE Trans. Neural Netw. Learn. Syst. 31(12) (2020) 5509-5521.


The goal of this Master thesis is to improve recent deep clustering models for HSI clustering. Following issues will be addressed:

1. Build an unsupervised neural network architecture, which allows simultaneous feature extraction and learning of cluster structure.

2. Introduce a spatial regularization to improve clustering results.

3. Develop an efficient algorithm to solve the resulting model and conduct experiments to validate the effectiveness of the developed model.

The students will start with the released codes of existing deep clustering models that are designed for computer vision. Real experimental data sets will be provided.